YannByron opened a new pull request, #9936: URL: https://github.com/apache/paimon/pull/9936
### Purpose Make RoboMIND AgileX ingestion and ACT training one reproducible benchmark workflow: `ingest → prepare → run → compare`. The importer now lives in `pypaimon.benchmark.act.robomind_agilex` and publishes four consistently named tables: `agilex__info`, `agilex__episode`, `agilex__frame`, and `agilex__stat`. - Define typed group/episode metadata and immutable episode, global-frame and episode-local frame indices. Document every field and its RoboMIND source mapping. - Keep original image payloads and state/action columns in the frame wide table; train directly through `ContiguousWindowDataset`, with anchor-only observations and a contiguous action horizon. - Store raw feature moments and the successful-train ACT normalization scope together in `stat`. Resolve all members through `info.tables` at one immutable Tag, publish info last, and reject incomplete groups or stale normalization snapshots. This changes the benchmark's sample import path, table names and metadata schemas. Re-ingest into a new warehouse and regenerate the experiment with `prepare`; existing warehouses and experiment files are not migrated. Publication uses ordered single-writer commits, not a cross-table transaction. ### Tests - `pytest -q pypaimon/tests/robomind_agilex_pipeline_test.py pypaimon/tests/act_runner_test.py pypaimon/tests/act_benchmark_test.py`: **58 passed, 1 skipped**. Includes local/Ray parity, exact four-table naming, Tag isolation, stale statistics, quality selection and ACT normalization scopes. The skipped case requires an explicitly supplied external dataset. - Flake8 on the changed benchmark/test files using `paimon-python/dev/cfg.ini`; `git diff --cached --check`. - Real RoboMIND AgileX smoke run: 4 HDF5 trajectories / 2,299 frames ingested with the combined stat schema. HDF5 and Paimon produced identical fingerprints for 59 windows and identical train/eval losses across three 20-step CPU ACT runs. This validates data/training parity, not policy convergence or production throughput. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
